AMD Confirms AI-Based FSR4 Will Boost Image Quality And Power Efficiency

AMD's FidelityFX Super Resolution (FSR) has been through a lot of iterations. The first major version was a straightforward spatial upscaler that produced unsatisfactory results unless the render resolution was very close to the output resolution. FSR2 was a major step forward, introducing motion vectors and temporal sampling to draw data from multiple frames, producing a significantly improved output image.

FSR3 introduced frame generation alongside improvements to the temporal upscaling of FSR2, but it was still entirely "analytical"—based on algorithms hand-tuned by humans. As most gamers who have seen both will tell you, NVIDIA's DLSS typically produces better results than even the latest FSR 3.1, and many credit this advantage to that technique's use of advanced visual AI running on the GeForce GPUs' "tensor cores".

Clearly, AMD's Jack Huynh agrees, because he recently told Tom's Hardware that the company's previous "analytical-based" methods were chosen for their "very fast time to market", while "the future is going ... AI based." Huynh is the SVP and GM of graphics at AMD; he specifically mentions FSR4 by name, and says that it is "going AI-based frame generation" with the specific goal of improving battery life for handheld gaming systems.

On the handheld side, my number one priority is battery life. If you look at the ASUS ROG Ally or the Lenovo Legion Go, it’s just that the battery life is not there. I need multiple hours. I need to play a Wukong for three hours, not 60 minutes. This is where frame generation and interpolation [come in], so this is the FSR4 that we're adding.

Because FSR2 and FSR3 were analytical based generation. It was filter based. Now, we did that because we wanted something with a very fast time to market. What I told the team was, "Guys, that's not where the future is going." So we completely pivoted the team about 9-12 months ago to go AI based.

So now we're going AI-based frame generation, frame interpolation, and the idea is increased efficiency to maximize battery life. And then we could lock the frames per second, maybe it's 30 frames per second, or 35. My number one goal right now is to maximize battery life. I think that's the biggest complaint. I read the returns too from the retailer, where people want to be able to play these games.

—Jack Huynh via Tom's Hardware

Huynh's quote was apparently in response to a question about the future of PC-based gaming handhelds. Huynh considers the biggest problem facing these devices to be battery life, and clearly sees frame generation as one possible solution to that issue. The idea is that generating interpolated frames is much less computationally-intensive than fully rendering out those frames, so you can lock the framerate at a relatively low number like the "30 or 35" that he suggests and then use frame generation to make it smooth, saving power.

This particular use case isn't ideal in our opinion, and all you need to do to prove this out is try it for yourself. Load up your favorite game with frame generation support, whether DLSS or FSR3, set a 60 FPS framerate cap, and then enable frame generation. What you get is a game that's laggier than 30 FPS with visible interpolation artifacts. Game Science did exactly this for the "Performance" mode in the PS5 version of Black Myth Wukong, and it's a misery to play in that mode.


However, FSR4 will presumably be available on other devices besides gaming handhelds, and it may well be a great feature for AMD's upcoming RDNA 4-based discrete GPUs. Huynh noted that AMD started working on an AI-powered frame generation solution "9-12 months ago", so FSR4 could feasibly be ready for the launch of AMD's new graphics cards, which is expected to happen around CES, in January—still about three months away.

There are still a lot of questions around FSR4, though. For starters, what AI hardware will it use? Will it take a page from Microsoft's AutoSR solution, and use the system's NPU to perform AI frame generation? Or will it require specialized AI hardware only available in AMD's RDNA 4 GPUs? Alternatively, it could use an AI model that can run on any CPU, GPU, or NPU—the platform-agnostic nature of FSR has been a hallmark of AMD's graphics software for a long time, after all.

More than that, we also wonder about the resolution upscaling portion of FSR4. In terms of image quality, AMD's frame generation solution is much more competitive with NVIDIA's DLSS versus its resolution upscaling technology. Even Intel's own XeSS upscaling often outperforms AMD's FSR resolution upscaling. Frame generation is neat and all, but we're much more interested to hear about AI-based resolution upscaling from AMD, especially if it's vendor-agnostic. Maybe we'll hear something at AMD's Advancing AI event on October 10th.